Before They Ask: Anticipating Help-Seeking from Behavioral Signals in Virtual Reality

Immersive virtual reality (VR) is increasingly used for training and has the ability to track learner behavior continuously. Yet support is usually delivered only once a learner asks for it, and help is not always sought when needed. While support based on performance is common, less is known about detecting an emerging need for help from non-verbal behavior. We examined whether an upcoming help request can be anticipated from movement and gaze alone. We analyzed data from 21 participants who completed a VR assembly task in which rules changed without warning and could be consulted through a deliberate, logged action. All models were evaluated on unseen participants. Trial-level prediction was accurate but relied mainly on task outcomes, such as errors and earlier consultations, rather than on the learner's ongoing behavior. Using short, continuously updated windows of movement and gaze, the model reached an AUC-PR of 0.159 against a chance level of 0.064, about 2.5 times the chance level, and most of its discriminative ability remained when the final turn toward the source of help was excluded. Before a request, movement slowed, gaze stayed closer to the head's forward direction, and fixations lengthened. Help-seeking is thus partly foreshadowed by behavior VR systems already record, though the signal is still weak when generalized to new learners, who varied substantially in how often they sought help.